如果 Agent 只是回答一个问题，答错了，用户可能再问一次；但如果 Agent 开始下单、参与调度、修改生产系统、调用真实业务能力，那么一次错误可能带来的就不只是体验问题，而是实际业务风险。那么，企业 Agent 的边界，到底在哪？
近日，InfoQ《极客有约》X QCon 直播栏目特别邀请 NoDesk AI CTO 王仿担任主持人，和飞猪算法专家吕德涛、货拉拉技术总监亓隆基、江苏移动 AI 方向 CTO 王上淇一起，共同探讨企业级 Agent 的自主权与生产实践。
部分精彩观点如下：
- 业务场景更需要确定性和可靠性，这是模型天然不具备的，所以我们得有一个明确的边界，让模型的不确定性在确定的业务边界内进行。
- 关键环节还是用规则的方式做，只是描述和生成规则用 AI，成本、边界、准确性都解决了，可解释、可撤回、可控也更强。
- 涉及用户钱包权益的一定放在人这边；自由度高的，比如生成回复、澄清模糊意图、处理长尾表达，可以交给模型
- 企业 Agent 的壁垒，不是生成一份看起来不错的答案，而是企业自己的数据、工具、算法、业务规则、评测体系在真实世界里能不能成立
12 月 18–19 日，AICon 全球人工智能开发与应用大会将在北京举办。本届大会以探索 A 应用边界为主题，设置 13 大技术专题，覆盖大模型后训练、推理优化、Agent 工程、知识工程、智能体评测与安全、自进化系统、AI 软件交付、原生架构及组织变革等核心方向，汇聚一线技术专家，深入分享 AI 技术突破与企业规模化落地的真实实践。
以下内容基于直播速记整理，经 InfoQ 删减。
为什么现在要讨论 Agent 的边界？
王仿：为什么模型能力越来越强，我们反而越来越需要讨论 Agent 的边界？
吕德涛：模型越强，能处理更长程的任务、理解更复杂的需求、接触更多的业务状态。以前答错一个问题，用户还可能再问一次；但在真实业务里，推荐的价格或库存错了，会直接导致创单或支付失败，一旦进入交易链路就是严重的体验损失。而且模型很容易给出一个看起来很靠谱、其实是编造的答案，比如编一个很符合业务规则的酒店房型 ID，C 端用户完全感受不到。所以边界不是模型强弱的问题，而是它强了之后，怎么让明确的事情百分百保证正确。
亓隆基：这两年模型从只能回答问题到能调用工具，但原理上它还是基于 token 预测的概率模型，一定存在幻觉，是一种不确定的能力。进入真实业务，下单推荐、联系司机、撮合决策都是一笔真实的交易，犯错影响就大。业务场景更需要确定性和可靠性，这是模型天然不具备的，所以我们得有一个明确的边界，让模型的不确定性在确定的业务边界内进行。
王上淇：原来是对话框，答错用户再问一遍；现在是业务系统，出问题了谁都不能接受，比如移动内部 OA 办公批文推荐，批错领导后果就很严重，所以应用层面要通过 Harness 对 Agent 能力做校验。另一个是我做平台更有体会的：完全交给 Agent 去调大模型，对算力和 token 是浪费，网关也未必承载得住。原来防的是 DDOS，现在还要防 Agent 的无意攻击、无意消耗。算力 token 都要花钱，重试次数和边界机制没控制好，toC 产品会被这种 Agent 攻击，toB 产品会造成任务拥挤，我们生产就出现过某个智能体出问题、把整个生产带崩的情况。所以边界一头是能力，一头是系统稳定性。
王仿：我们有个客户是国内服饰领域一个 top 企业，内部 Agent 没做限制，直接去调内部数据库接口、并发也没限，把接口打崩了，影响线上半个多小时。所以从稳定性到业务逻辑，都要保证生产的确定性：服务得稳，业务规则在合规前提下执行，而不是跳出边界做事。
今天很多 Agent 已经能查信息、做分析、给建议，几位的实践却都在往更深处走，开始调用系统、改变状态、参与决策。Agent 从聊天助手到真正执行任务，跨过的是什么门槛？
吕德涛：跨过的不是会不会调用工具，而是能不能对最终结果负责。比如帮用户订上海迪士尼两大一小的门票：传统助手给个建议就好，到了业务系统里，要判断票种是不是成人票、两个大人一个小孩、人群范围对不对、有没有库存、组合出来的价格是不是最低，再进入订单流程，错一项都是回答得不好。我们会做几层约束：用模型做需求理解，票的组合、总价计算放到求解器里，保证结果可验证、可复算；酒店 ID、房型 ID 这类业务 ID 不会把原始值丢给模型，而是用短 ID 替换，展示时再回溯复原；涉及用户钱包的权益和支付，不直接开放给模型，要用户再确认。所以真正要跨的是校验、权限控制、可追溯的闭环。
亓隆基：我理解真正跨过去的，是原来很多必须由人做的事情，第一次可以由 AI 系统来完成。用户下单第一步是选车，我们有小车、大车，一个要运双开门冰箱的用户该选什么车？有了 Agent，我们把这些原本要用户自己具备的物流知识灌进去，让它帮用户理解意图、选车。这样业务系统就从单纯执行，变成了具备意图理解的能力。
王上淇：两个坎。第一是用户对 AI 的接受度不高。以前很多 AI 产品都是秒出的，而大模型要把文字转成向量、调接口、返回再渲染成 UI，流程完全不一样。我们给客户经理生成拜访方案，以前靠模板检索，现在让它先猜客户经理明天可能拜访哪家公司，提前结合内部资料和企查查把文档生成好，尽量命中请求，现场生成也能快很多。所以很大一部分工作是设计应用架构，该用 AI 时用 AI、不该用时少用，提前做缓存、预执行、预匹配。
第二是权限。企业内部能力原来都靠人操作，责任在人；toC 推荐对错用户还无所谓，企业内部批错单、批了不该批的，甚至没法追责。所以我们花很大精力做智能体权限管控，包括接口权限、MCP 能力权限，有权限的人才能用，责任就能落到人；遇到责任分摊和权限确认，就用 human in the loop 跟人确认，按公司规章设计 AI 交互。这有点像自动驾驶，L4 一直推不出来，L3 卡的不是技术，是职责和权限划分。
王仿：行业里不少 Agent 目前仍停留在演示或小范围试点阶段。toB 一个很重要的挑战是准确率，大家很少谈 Agent 准确率，默认模型聪明就不是问题，其实不是。比如进出口报关，前线业务同学拍的各种商品图片什么样的都有，先要解决感知问题，把不同模态的输入变成结构化文件，再审核、查询、交叉确认，生成报关和合同文件，以前全靠人做，一个报关要现场调研、四五十分钟以上。Agent 在这里最大的问题是准确率保证不了，不敢落地：手拍图片不同角度，解析商品编码、时间、日期戳很难保证；报关又很严肃，一个字段错整个申请就可能被驳回。
另一个影响准确的问题，是真要和业务专家面对面，把业务的目标、边界、拆解逻辑、背后依赖的规则都萃取成 Agent 能懂的语言，我们叫业务翻译。toB 和 toC 不一样，是人做了 80% 的活交给 Agent，而且一来一大批，十几个、几十个步骤走完还要稳定在 90% 或 95% 以上。除了权限，还有过程不能黑盒：人可以不看，但想看哪个环节，点进来就要把 Agent 的全过程，解析了哪些字段、为什么通过、为什么不通过、背后依赖什么逻辑，都提示出来。QClaw 刚出来时我以为企业端落地能进一大步，但并没有，因为它黑盒、保证不了准确，也解决不了权限。这在 toB 非常显著，也是 Agent 大规模落地的阻力。
吕德涛：我们有类似的场景，toB 客服的退款、退改规则，错一个就是一笔资损。
王仿：我见过一个做宠粮的企业做包材审核，包材审核出现疏漏，价值几百万商品就没法售卖。我原以为包材审核很简单，真做起来不是：要审的字段有 60 到 70 项，每项背后是不同国家的政策法规，不同食品有不同的要求。最近一个客户做审核类 Agent，一个品类的审核就有 200 多项需要审核校验的字段，商家要提交六七十份文件，每份提取几百个字段，有些规则能直接判断，还有很多要多跳，ABC 一套组合、ABD 又是另一套，业务专家要梳理好几天才慢慢的齐备。这就是 Agent 到复杂场景落地要跨的一关，背后专业性很强。技术同学做完这样的场景，基本就成了这个领域半个业务专家，从整体逻辑到细节都得懂，才能把它变成 Agent 能力；不懂、边界不清，做出来就上不了线，都是 demo，一遇异常就失效，业务上完全不能接受。
亓隆基：准确率问题，当前只靠大模型解决不了。比如选车，怎么 100% 按业务规则选对：拉冰箱必须直立，因为压缩机，但按长宽高算横着也能放；异形件的摆放规则更多。AI 运营里给运营用的 Agent 要生成营销、价格配置，多是精确计算，只能用工程方式。外呼聊价格、提取价格也可能出错，比如该加 10 块加成 100，在线上就是事故。所以这些都得用 Harness 的方式处理，或者限制住，把确定和不确定的地方定好边界，才可能在线上产生价值。
Agent 的边界到底怎么定？
王仿：面对某个真实业务任务，你们依据什么决定，这一步交给 Agent，那一步交给代码或传统算法，还有一步必须由人处理？
王上淇：分工上还有个现实问题：你做很多工程化，leader 会问你创新点在哪，都上大模型了，为什么流程引擎还是规则？刚才冰箱那个在我看来也是规则。但模型能力其实已经通往 AGI，最新出的国外模型生成代码很牛，因为代码能自闭环；企业里不好，是因为企业权限、企业信息、已知知识模型不知道，比如冰箱必须竖着、审核什么情况 AB、什么情况 CD，这类规则通用模型永远学不到，企业又不太可能微调，我们江苏移动就没有任何微调，让程序员做也难，因为程序员不懂业务。
所以我们现在的做法是：让业务人员通过 AI coding，把确定的、需要规则的业务规则固化成代码，主流程仍然用 AI 串。比如内部派单流程，原来是纯规则引擎编排，现在让业务专家描述规则，AI 生成一段代码，业务人员不用校验代码本身，流程再走时调用这段代码。出了错就判断：是意图识别错了，就去改 Harness 的提示词或补条件；是规则描述对、但生成的代码不对，就让业务部门给测试用例校验后改代码。一轮轮迭代下来，关键环节还是用规则的方式做，只是描述和生成规则用 AI，成本、边界、准确性都解决了，可解释、可撤回、可控也更强。
亓隆基：我们原来也很苦恼哪些用 Agent、哪些用工程。比如用户助手、司机助手，原来在图形界面上点，改成对话方式后，发现在老客身上是左手倒右手，效率反而降低。所以我们认定，Agent 应该用来拓展系统能力，而不是把已有能力重做一遍：把原来做不到的、需要根据当前状态重新思考规划的事，以及萃取出来的业务知识，交给 Agent；而实时交易、规则、传统算法（比如接单概率预测）这些有解的、确定性的问题，还是用最擅长的工具。Agent 更多是把现有能力组合串联，实时思考规划，为业务目标持续努力。
比如 AI 撮合解决的是长尾订单：原来播单给司机，司机十分钟、二十分钟不接，就按规则再播一轮；现在可以根据用户、订单的状态实时规划，用 AI 外呼跟用户、司机商量价格、问他为什么不想接、多少钱能走，收集到足够多司机的意愿再去促成这单。这在没有 Agent 之前只能靠客服或专职人工完成。总结就是，确定有解的问题用工程或机器学习，开放性、需要持续规划决策执行反馈的问题用 Agent。我们做了很多 Agent、花了很多 token，也一直在想额外的业务收益到底是什么，token 账单每月都真实存在，ROI 到底正不正，其实很难回答。所以要把 Agent 用在真正能带来业务收益、提供额外价值的地方。
吕德涛：跟两位不太一样，我们主要是 toC。涉及用户钱包权益的一定放在人这边；自由度高的，比如生成回复、澄清模糊意图、处理长尾表达，可以交给模型。就拿组票来说，最初我们也是用当时最好的模型去组几大几小的门票，本以为没问题，但准确率也就百分之八十左右，最后还是变成模型提需求、用传统的混合整数规划来求解。
王仿：这其实是两套范式、两套技术栈，和深度学习差别很大，适用场景也小。传统推荐系统主要依赖用户行为和交互日志，通过协同过滤、矩阵分解或深度模型优化预测目标；企业 Agent 除了从数据中学习，还要把业务规则、专家知识和工具执行链路显式组织起来。两者面对的问题和工程约束并不完全相同。
我还参观过制造业，他们换耗材，像果链、手机制造里的精密仪器，原来用满一万次就批量换掉，挺浪费；现在结合 Agent 判断健康度，其中传统深度学习负责感知，从图像、数值去判断健康度的波动。以前是人来判断该不该换、是误报还是真的，现在用 Agent 代替人做决策，人工和损耗都降下来了。所以哪些用人、哪些用传统算法、哪些用 Agent，其实没有统一范式，还是要结合具体业务。
那做这种分工设计，什么样的人合适？懂业务的、懂技术的，还是都要懂一些？
亓隆基：前几天某大厂发了一篇文章，说花了无数 token 得出一个关键认知：AI 工程是业产研的工程，不能只靠产研。我们的经验也是这样，真要把业务 Agent 做好，业务、产品、研发都得在，因为只靠产研，你不知道业务策略是什么。就拿 AI 撮合说，怎么跟用户、跟司机打电话，这些策略只有真实客服和一线业务最清楚。所以业产研都要在，而且对技术来说，这比原来做工程项目更有挑战：业务会觉得你做的是最先进的生产力，就该告诉我怎么做最好，而实际上应该是业务和产研一起，把有效的业务策略萃取出来，再让 Agent 遵循执行、拿到结果。
王仿：再把问题收窄。我们常用准确率、置信度来判断能不能上线，但真实业务里有另一个更直接的标准：做错了能不能撤回。可撤回性是不是一个足够通用的边界判据？如果不是，还要加上哪些因素，比如虽然能撤回但恢复成本很高，或者概率很低但后果特别严重？
王上淇：可撤回我们是有要求的，但它不是充分必要条件。尤其是订单，我可以用 AI 下订单，但不能把原来的下单、取消界面删掉，因为再回头跟 AI 说取消已经来不及了。所以现在所有 AI Agent 原有的流程都还在，撤回其实是靠原有 GUI 流程保证的。还有恢复成本，它甚至决定允不允许 AI 操作：比如盖章、审核自动化，本身不复杂，但涉及金额或责任特别重大时，业务人员会觉得哪怕 99% 准确也不让做，恢复成本接受不了。另外可解释、可复现也很重要。深度学习对可解释性要求不高，看指标、看数据就行；但在生产场景里 Agent 是代替人操作的，业务人员不知道它在干什么就接受不了。要能像 Claude 那样一步步告诉你在干什么，不对可以立刻暂停，可暂停、可解释很重要。
亓隆基：还是要看影响面能不能接受。比如 AI 运营改后台的营销、价格配置，即使能撤回，一旦上线就会影响非常多订单，影响不可接受就得在上线前加一道人工审核。我们默认 Agent 保证不了 100% 准确，关键操作上接受不了后果，就必须人工确认。AI 外呼每个版本上线都会做红线检测，看会不会说错话。有些场景可以放松，比如提取金额出错，就通过评测观测资损率，在可控范围内就允许上线。所以最后还是看犯了错、产生的影响能不能接受，能接受多大范围、多大资损，据此决定放行还是人工确认。
吕德涛：我们跟亓老师很像。比如订单日期推荐错，用户要周五、我们推了周六，日期在导购卡片上又不显眼，用户可能直接就付款、还用了优惠券，这时再回撤，权益已经没了，酒店和交通的价格变价又快，好价格就没了，所以特别小心。第二还要看错误用户会不会发现，比如行程规划明显不合理、来回往返同一个景点，或者推荐理由是 A 酒店却挂了 B 酒店的卡，这些都不行。
王仿：所以 toC 更关注用户体验，toB 把安全和责任作为红线，报关出错到底谁扛，Agent 出错会不会影响我的绩效。
观众：Agent 在业务场景里怎么自我优化迭代？
王上淇：分两种。业务场景的优化迭代，现在是找 badcase 分析，判断到底是意图识别错了、工具选错了，还是工具选对了但效果或格式不符预期：意图识别错就改提示词或换模型，工具调用错就评估是描述不准还是工具太多要删减，说白了就是要么改、要么换模型、要么改 Harness 的配套。另一块我们做 MAS 平台、AI 观测和 AI 稳定性，还在探索。现在 70% 的智能体优化还是人工的，业界有一些偏自动的方式，通过 trace 去优化 harness 或 skills，说实话在生产场景效果不好、也不可控，但在通用场景还可以。
所谓通用场景，就是比如 OpenClaw 很火、移动也做一个内部版本给员工用，各部门搞出各种 skills，处理文档、处理投诉。之前拿开源的 Harness 改了改准确率不高，领导要求准确率、不能有投诉、还得有评估标准，我们就想用调用链来评估，比如超过 20 步的技能，让大模型根据 trace 自动分析是 skill 生成得不好还是提示词不好，再动态优化。这种偏泛化的通用场景里有用，能把步数从普遍超过 20 步、耗时十几分钟，平均降到 15 步以内，大概提升四分之一效率；准确性不好评估，就看重复生成率有没有下降。所以专一的业务场景还是靠人评估、靠测试集，通用场景可以试 trace 自动化。
亓隆基：我们的经验是有两个闭环。外环是业务指标的 loop。比如 AI 撮合解决订单配对，环节很多、各有转化漏斗，得先有业务指标体系，确保在解决正确的问题：先找到正确的司机，再用 AI 外呼，靠话术 SOP、指令遵循、Agent 能力去跟用户聊，最后成单；中间还有外呼线路的触达率、接通率。外环就是确保能找到掉点的环节和正确的做功方向，接通率上不去，Agent 再强也没收益。内环是 Agent 本身能力的闭环。AI 外呼是开放式的，就要靠可观测性记录全过程，看每个环节的意图识别、ASR、回复是否正确，很多时候要人工标注，业内会用 LLM as a Judge 加速，也可以用更高级的模型来标注，省掉大量人力。另外还可以用大模型去萃取金牌销冠的话术和策略，加速业务知识萃取。所以要迭代 Agent，两个闭环都得做好：业务指标体系定好，明确 Agent 在哪个方向做功；Agent 自身能力上，话术和业务策略有没有被遵循、badcase 率有没有降下来。
吕德涛：一是评测一定要做准，我们线上入库标准一部分靠强代码校验、一部分靠价值模型，评测不准，自进化的方向就错了，所以要在评测上下功夫。二是准备一份线上的黄金数据集，叫 metadata，每次迭代都在上面做回测，保证修改是准的，因为自进化很容易 hacking，模型很容易偷懒。
王仿：核心是要有一个可学习的目标。亓老师讲的，我得知道我学的是金牌客服、金牌销售的逻辑，这个 tag 在哪；还得圈定这个 tag 下是什么场景、用什么策略。自进化就得有这么一套反馈逻辑，包括吕老师说的拿金标准做离线回测。先定义好问题，语料丰富了，Agent 的自我学习反馈能力会越来越强，再和线上指标挂钩。我过去在客服场景这么做过，很有效：离线挖掘能发现，金牌客服在人问优惠条件时，会讲怎么用优惠、再用产品卖点击中他，影响他的决策；因为你已经有大量转化和未转化的结果，未转化的往往是人问啥答啥、很泛、转化率很低。
Agent 应该拥有多少自主权？
王仿：现在行业里很喜欢谈 Autonomous Agent，好像越能自主规划、调用工具、完成长任务就越先进。但到了生产环境，真实业务里的 Agent 真的是越自主越好吗？
亓隆基：我有两个反例。第一个是下单，我们试过完全 agentic，把上下文、目标、可调用的工具都给模型，让它自主决定怎么跟用户交互，结果回收回来比原来定好的流程还差。因为在明确业务场景下，AI 会输出非常多内容，提问只花几秒，读懂却要很久；自主时它给用户太多选择，但明确的用户只想省心，你直接告诉他选什么车、什么价就行。你自己用 AI 也这样，做 PPT 只想让它快点做好，不想讨论半天。第二个是 AI 调度，同样是 agentic，把目标、上下文、工具都交给它自己决定怎么跟用户和司机聊，问题是路径不可控：最后有效了，也不知道它用的什么策略，就没法持续优化，来了新策略也没法让它遵循去回收效果，从策略到执行到结果的关系被打断了。所以关键节点需要确定性，只在需要模型自主发挥的环节让它自主。
吕德涛：生产环境里 Agent 要有明确的界限和责任，即使没到 100%、也要 99% 以上才行，否则对 C 端就是很大的伤害。它可以在决策空间里灵活，比如调用各种工具；但输出给 C 端的回复要谨慎，组票、实时价格、库存、商品 ID、订单状态都不能错，错了不是舆情就是资损。
王上淇：换个角度想，自主和失控的区别是什么？我理解自主是在有限规则下发挥最大潜力，失控是不按规则来。我们为什么要聊自主性？如果任务本来就很 work，那也需要 Agent，但绝不需要它越来越自主；自主性是用来解决原来解决不了的问题的。所以自主性的上限，就是你希望它解决问题的灵活度，但要有门槛，没门槛就是失控。门槛从技术上很简单：允许你调用工具，但能调哪些由我定，你的自主权只在于自己选调哪个，门槛就是我配 10 个还是 20 个工具。
这里还延伸出一个问题，原来的编程叫可信编程，现在通用编程叫不可信编程，因为它会写代码：给了它 10 个工具，它觉得完不成但还是硬做，就自己写代码去实现，事情就又不可控了，所以大家都要做沙箱。另外，为什么自主不够好？因为 99% 的厂商没有能力基于它的行为或错误去提升：agentic 的提升方案很少、harness 又复杂，甚至找不到问题在哪，因为大部分业务场景没法把数据收回来做强化学习，成本太高。未来如果有企业能这么做，自主性才可能越来越高，否则一定被限制在门槛里，不然就是失控。
王仿：相比推荐、广告等已经形成成熟反馈闭环的系统，企业 Agent 的优化链路还不够标准化。虽然已经可以通过轨迹分析、离线评测和反馈数据迭代，但要做到可归因、可控风险地持续提升，目前仍然很难，也缺少可广泛复用的生产范式。
展望未来
王仿：未来真正有价值的企业 Agent，核心竞争力到底是什么？
亓隆基：模型是起点，基座模型大家都能选，选个好模型起点就高一些。到了业务上，第一步是 Harness，用工程把确定和不确定的边界分清：需要思考规划的交给模型，需要确定的交给工程，再把机器学习算法、运筹优化这些工具综合进来，这是企业 Agent 的下限。下限之上，靠业务 loop 和 Agent 能力 loop 这两个闭环持续提升。上限则像做工程一样，一个业务系统好不好，最后不在于基础设施，而在于对业务知识的萃取和业务架构的设计，它决定了系统能支持业务五到十年，还是每年都要大重构。类比到 Agent，上限就在于能萃取多少业务知识、找到什么场景和做功方向。所以就是模型决定起点，Harness 决定下限，双 Loop 决定方向，加上业务知识萃取到什么程度、在哪个环节做功，决定了上限。
王上淇：企业 Agent 是我自己打造一个，还是我帮别的企业打造，核心竞争力和壁垒是不一样的。对我们来说，怎么快速把企业能力对接上就是个大困难，我们做过 MCP、Skills（可能还不算协议类型）、AGUI 这些规范，但把存量系统接进 Agent 都很难。数据萃取这点我非常赞同，没有数据、没有数据萃取的企业 Agent 就没有意义。企业生产场景最大的问题是，规则和业务数据大多沉淀在原有系统里。像移动，还有中石油等等，我们都做过一些服务，你想给企业做诊断会发现很难，因为拿不到、也接不上它的系统。所以怎么用相对通用的方式，把信息化不好的企业的能力梳理、暴露出来，变成一个能快速接入的平台，是未来的关键。
之前很火的 Palantir 做数据业务建模，很早就有了，但能力释放不出来，梳理好了数据、有了平台，也用不起来，直到有了大模型才有人能去用。现在大部分企业缺的就是这样一个 Palantir 式的平台，能把数据资产自动化地萃取出来。另外还有一点，企业内部关键是怎么平衡管理和智能体本身：企业 Agent 能不能承载职责和流程。我们现在不只做业务 Agent，还把智能体和原有业务流程挂钩，流程对应原来的责任，比如智能批文、审核都要求智能体跟流程对应，才能保证责任和权限清晰。所以一方面是萃取企业能力数据的能力，一方面是把 Agent 和流程关联的能力，都建立在足够强的 Harness、loop 和数据能力之上。
吕德涛：旅行是一个相对垂类的行业。我觉得未来企业 Agent 的核心竞争力，还是垂类领域的行业供给和行业知识，能不能变成可执行、可验证、可迭代的系统能力。比如行程规划，调基模就能产生一份不错的行程，但难的是它能不能真的走得通：景点之间的距离、交通、营业时间能不能衔接上，还要考虑每天游玩的强度，景点在现实中得真实存在。行程往下就是买酒店、买门票，价格库存必须来自真实业务系统，组出来的票要符合真实出行需求，进订单还要用户确认，这些都不是模型自由发挥能生成的。模型适合理解偏好、筛选景点、生成方案，地图路线营业时间可以用工具取，票的组合可以用确定性算法，但价格和库存必须由业务系统提供。所以企业 Agent 的壁垒，不是生成一份看起来不错的答案，而是企业自己的数据、工具、算法、业务规则、评测体系在真实世界里能不能成立。
王仿：这有点像世界模型，在真实世界里到底 work 不 work。过去两年我们一直在讨论怎么让 Agent 更聪明、做更多事；但 Agent 真正进生产之后，另一个问题越来越重要：我们到底允许 Agent 做什么，哪些必须由人来做？
热度延续：12 月 AICon 北京站见
12 月 18–19 日，AICon 全球人工智能开发与应用大会将在北京举办。本届大会以探索 A 应用边界为主题，设置 13 大技术专题，覆盖大模型后训练、推理优化、Agent 工程、知识工程、智能体评测与安全、自进化系统、AI 软件交付、原生架构及组织变革等核心方向，汇聚一线技术专家，深入分享 AI 技术突破与企业规模化落地的真实实践。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
If Agent answers only one question, wrong answer, the user may ask again; but if Agent starts to order, participate in the movement, modify the production system, and call for real business capability, then an error may bring more than an experience, but an actual business risk. So, where is Enterprise Agent's border?
In recent days, the InfoQ Telegraph X QCon live section specifically invited NoDesk AI CTO to act as moderator and to discuss enterprise-level Agent autonomy and production practices, together with the expert in flying pig algorithms, Ludtao, the technical director of Salala, Yongji, and the CTO King, in Jiangsu ' s Mobile AI direction.
Some of the good points are as follows:
- The business landscape requires greater certainty and reliability, which the models are naturally not available, so we need to have a clear boundary that allows the uncertainty of the model to take place within the defined operational boundaries.
- The key links are also done in a rule-based manner, with only AI describing and generating the rules, with costs, boundaries, accuracy all resolved, interpretable, reversible, manageable.
- The rights of the user's wallet must be on the people's side; the freest ones, such as generating responses, clarifying vague intentions, processing long-tail expressions, can be given to models
- The barriers of Enterprise Agent, not to produce a good answer, but to be able to set up its own data, tools, algorithms, business rules, evaluation systems in the real world.
On December 18-19, the Aicon Global Congress on Artificial Intelligence Development and Applications will be held in Beijing. The Congress, which focused on exploring the A application boundary, set up 13 major technical topics covering the core directions of LLM post-training, reasoning optimization, Agent engineering, intellectual engineering, AI agent assessment and security, self-evolution systems, AI software delivery, native architecture and organizational change, bringing together first-line technical experts to share in depth the real practices of AI technological breakthroughs and enterprise size.
The following are based on live shorthands, which are deleted by InfoQ.
Why discuss the boundaries of Agent now?
Wang Moo: Why are models getting stronger and we getting to discuss the boundaries of Agent?
Ludtau: The stronger the model, the better it can handle longer-term tasks, understand more complex needs and engage more business. Users may have asked the wrong question once before; but in real business, the recommended price or stock is wrong, leading directly to a single or payment failure, which is a serious loss of experience once entered the trading chain. And it's easy for a model to give a pretty good, actually a fictional answer, like making up a hotel-type ID that fits the rules of business. So the border is not a question of a model that is strong and weak, but of how it can be made absolutely right when it is strong.
亓Longki: The two-year model ranges from answering only questions to being able to call tools, but in principle it is based on a probabilistic model of token's prediction, there must be an illusion and an uncertain ability. In order to enter real business, the following recommendations, contacts with drivers and concerted decision-making are real transactions and the impact is greater by mistake. Business scenes require greater certainty and reliability, which the models do not naturally have, so we need to have a clear boundary within which the uncertainty of the models can take place.
Q: It's a dialogue box, and the wrong user asks again; it's a business system, and it's not acceptable to anyone, such as moving internal OA office approvals, and the consequences of wrong leadership are serious, so the application level needs to be validated by Harnesse's Agent capabilities. And the other one is more familiar with my platform: it's completely assigned to Agent to tune LLM, which is wasteful for math and token, and the gateway is not necessarily loaded. It was DDOS, and now it's Agent's unintentional attack, unintentional consumption. Token spends money, retry times and border mechanisms are not controlled, toc products are attacked, tob products are overloaded, and we have a problem with AI agent, and the whole production has collapsed. So the boundary is capacity, and the border is systemic stability.
Wang Moo: One of our clients is a top company in the country's garment industry, and there's no limit to the inside, and there's no limit to just dialing up the internal database interface, breaking it down, affecting the line for more than half an hour. Thus, from stability to business logic, there is a need to ensure certainty in production: services are stable and operational rules are enforced in compliance rather than out of the border.
Today, many Agents have been able to access information, analyse and advise, while several practices have gone deeper, starting to call systems, change state, and participate in decision-making. Agent, what's the threshold between a chat assistant and a real mission?
Ludtau: It's not about calling tools, it's about being responsible for the end result. For example, the user is given two small tickets to Shanghai Disney: the traditional assistant gives a suggestion, and in the business system, the question is whether the vote is adult, two adults are a child, the population is not in the right range, there is no inventory, the price is lowest, the order process is followed by the wrong answer. We have several layers of constraint: modeling for demand understanding, combinations of tickets, total price calculations placed in solvers to ensure that results can be validated and recalculated; hotel ID, room ID, such business IDs do not throw original values into models, but replace them with short IDs, with retroactive recovery on display; interests and payments involving user wallets are not directly open to models and users are required to confirm them. So what really needs to be crossed is verification, permission control, retroactive closure.
Rongji: I understand that there are many things that have to be done in the past that can be done by the AI system for the first time. The first step for a user is to choose a car. We have a car, a big car, a user who's going to double-open the fridge. With Agent, we're going to fill in the logistics knowledge that the intended users have, so that they can understand the intentions and choose the cars. Thus, the operational system has been transformed from a simple implementation to an understanding of intent.
The King: Two. The first is the low user acceptance of AI. Many of the AI products used to come in seconds, and LLM wanted to convert text into vectors, adjust interfaces, and return to render UIs. We have created a visiting programme for client managers, which was previously searchable by template, and now allows them to guess which company the customer manager might visit tomorrow, to generate the file in advance, by combining in-house information and by searching, and to do as much as possible to hit the request, and to do so much faster. So a large part of the job is to design the application architecture, using AI, not using less time, cache in advance, pre-implementation, pre-matching.
The second is permission. The enterprise's internal capacity was originally people-operated and the responsibility lay with them; the ToC recommendations were not relevant to the wrong user, and the enterprise had been wrongly approved, wrongly approved and could not even be held accountable. So we spent a lot of effort on the control of AI agentpermission, including the interface permission, the MCP capability permission, the ability of the permission to use it, the responsibility of the permission, and the burden sharing and confirmation of permission, the use of humans in the loop, the design of AI by company regulations. It's kind of like auto-drive, L4 can't get out, L3 card's not technical, it's permission.
Wang Moo: A lot of Agent in the industry is still in the demonstration or small pilot phase. One of the important challenges is accuracy, and there's little talk about Agent accuracy, and defaulting on smart models is not a problem, actually. For example, import and export declarations, all kinds of images of commodities taken by front-line business students address the problem of perception, transform the input of different models into structured documents, then review, query, cross-check, generate customs and contract documents, which were previously done by people, and a declaration requires on-site research for over 40 minutes. Agent's biggest problem here is that the accuracy rate is not guaranteed, and it is not safe to drop: the analysis of the commodity code, the time, the date stamp is difficult to secure from different angles; the declaration is serious, and the entire application of field may be rejected.
Another issue that has an impact on accuracy is the face-to-face presence of real and operational experts, drawing from the objectives of the operation, boundaries, dismantling logic, and the rules relied upon behind it the language that Agent understands, which we call business translation. Unlike toc, people do 80% of their work to Agent, and a lot of, a dozen, dozens of steps are going to stabilize at 90% or more. Apart from permission, there are processes that cannot be blacked out: one can look away, but if one wants to look at one, one has to point in the whole Agent process, decipher which field was passed, why not, why not, and what logic lies behind it. When QClaw first came out, I thought the company could make a big step forward, but it didn't, because it was black box, guaranteed no accuracy, and could not solve permission. It's very visible in tob, and it's also the resistance of Agent to mass landings.
Ludtow: We have a similar scene, a refund from Tob's customer service, a refund rule, and a loss.
Wang Mo: I've seen a food-based company do a package audit, and it's leaking, and millions of goods can't be sold. I thought the package audit was simple, and it didn't really happen: field was about 60 to 70, each with different national policies and regulations and different food requirements. Agent, a recent client's audit class, has over 200 fields for verification, 60-70 documents for merchants, several hundred fields for each, some rules for direct judgement, many more jumps, ABC combinations, ABDs, and a few days for business experts. And that's what Agent's going to do to the complex scene, which is very professional behind it. After doing this, tech students become basically half-experts in the field, from the whole logic to the details, in order to turn it into the Agent's ability to understand, undefined borders, unable to reach the line, demo, out of the ordinary, completely unacceptable business.
Rongji: The problem of accuracy rates cannot be solved by LLM. For example, car selection, how is 100% paired by business rules: the refrigerator has to be straight because of the compressor, but can be put at long width; there are more rules for the placement of alien objects. AI Agent, which operates for operations, produces marketing, price configurations, many of which are precise calculations, only engineering. There may also be errors in talking about prices and extracting prices, such as the addition of $10 plus 100, which is an accident online. So all of this has to be handled in a Harnessy way, or limited, to establish boundaries where there is certainty and uncertainty, so that it can produce value online.
What about the boundaries of Agent?
Wang Moo: On what basis do you decide that this step should be given to Agent, that it should be given to code or traditional algorithms, and that one must be handled by someone else?
Q: There's a real problem with the division of labour: you do a lot of engineering, leader asks you where you're going, you're on LLM, why the process engine or the rules? The one in the fridge was the rule to me. But the power of the model actually leads to AGI, where the latest foreign model generation code is very good because the code is self-contained; the business is bad because the company permission, the business information, the known knowledge model does not know, for example, what the refrigerator has to stand up, what the AB, what the CD, the generic model of such rules will never learn, and it is unlikely that the company will fine-tune it, and it is difficult for the programmer to do it, because the programmer does not understand the business.
So what we're doing now is getting business people to use AI coding to consolidate established, rule-based business rules into codes, and the main process still uses AI strings. For example, the internal tasking process, which was originally organized as a purely rule engine, now allows business specialists to describe the rules, AI to generate a code, business people to use the code when the process moves again without checking the code itself. Errors are judged: the intent is to identify wrongly, to change Harness ' s prompt or condition; the rules are correct, but the resulting code is incorrect, and business departments are asked to verify and change the code for testing. A round of rotations, where key links are done in a rule-based manner, only describes and generates the rules in AI, where costs, boundaries, accuracy are resolved and can be explained, withdrawn, controlled.
Rongji: We used to worry about the use of Agent and the use of engineering. For example, the assistant to the user, the assistant to the driver, who had used the graphic interface and had changed it to the mode of dialogue, found that it was the left hand to the right hand on the elderly, which was less efficient. So we're thinking that Agent should be used to expand the system's capabilities, rather than re-doing what it is already capable of: handing over to Agent what was not possible, what needed to be rethought in the current state, and what was derived from business knowledge; and real-time transactions, rules, traditional algorithms (e.g. billing probability predictions) are solving, definitive questions, or using the best tools. Agent is more of a combination of existing capacities and real-time thinking about planning and ongoing efforts towards operational objectives.
For example, AI set up a long-tail order: the original flyer was sent to the driver, who did not pick up for 10 minutes or 20 minutes, and the rule was followed; it is now possible to plan in real time the user, the order's status, using an AI page to negotiate the price with the user, the driver, ask him why he didn't want to pick up, how much money he could leave, collect enough of the driver's will to make it. This can only be done on a passenger service or a full-time manual basis without Agent. In conclusion, the problem of identifying solvency issues for engineering or machine learning, openness, and the need for continuous planning for feedback on implementation of decision-making, Agent. We've done a lot of Agent, we've spent a lot of token, and we've been thinking about what the extra business proceeds are, and token's bills are real every month. So we need to use Agent where it actually brings business returns and extra value.
Ludeto: Unlike you two, we are primarily toc. User wallet interests must be placed on the side of the person; those with high freedom, such as generating responses, clarifying vague intentions and processing long-tail expressions, can be handed over to models. In terms of ticket numbers, we used the best model at the time to form a few large and small tickets, which we thought was fine, but the accuracy rate was about 80 per cent, and it turned into model demand and traditional mixed integer planning.
Wang Moo: This is actually a two-set paradigm, two-set technocratic collection, and a very different depth of learning, and the applicable scenario is small. Traditional referral systems rely mainly on user behaviour and interactive logs to optimize predictions through synergistic filtering, matrix decomposition or in-depth modelling; in addition to learning from data, Enterprise Agent organizes business rules, expert knowledge and tools to implement a chain-of-visibility approach. The problems and constraints faced by both are not identical.
I also visited the manufacturing industry, where they changed materials, like the precision instruments in fruit chains and mobile phones, which had been replaced in 10,000 batches, which was a waste; and now, in conjunction with Agent, the assessment of health, in which traditional depths of learning are responsible for perception, and health fluctuations are judged by images, values. People used to judge whether or not to change, falsely report or really, and now people are taking decisions instead of Agent. So who uses people, who uses traditional algorithms and who uses Agent does not have a uniform paradigm or does it combine specific operations.
So what's the right person for this division of labour? You know business, you know technology or something?
亓Longki: A few days ago, a large factory published an article saying that it took countless tokens to come up with a key idea: AI works as an industrial research project and cannot rely on production research alone. And that's what we've been doing, and we really have to do business, business, product, research, because you don't know what business strategy is. Take AI and tell them how to call the users and the drivers, and these tactics are best known only in real customer service and first line business. So it's going to be more challenging for both industry and technology than it was for engineering: business thinks you're doing the most productive work, and you should tell me what's best, and it's actually going to be doing business and research together, extracting an effective business strategy and getting Agent to follow implementation and get results.
Q: Let's narrow down the problem. We often use accuracy and confidence to judge whether or not to go online, but there is another more direct criterion in the real business: whether to do wrong or not. Is revocability a sufficiently common border test? If not, what additional factors should be added, such as the high cost of recovery, if possible, or the low probability of recovery, but the consequences are particularly severe?
Wang Jing: There are demands to withdraw, but it is not a sufficient necessity. In particular, I can place an order with AI, but I can't delete the original listing, cancel the interface, because it's too late to talk to AI about cancellation. So now all AI Agent's original processes are still there, and the withdrawal is actually guaranteed by the original GUI process. There are recovery costs, and it even decides not to allow AI operations: stamping, automating audits, for example, are not complicated per se, but when the amounts or responsibilities involved are particularly significant, the operational staff feel that even 99 per cent of the costs of recovery are inaccurate and unacceptable. It is also important that it be explained and reproduced. In-depth learning is not so demanding for interpretability as to look at indicators and data; but in the production scene, Agent is a substitute for people who cannot accept it without knowing what it is doing. It's important to be able to tell you one step at a time like Claude, not to suspend immediately.
亓: It depends on whether the influence is acceptable. For example, AI, which operates to change backstage marketing, price configurations, even if withdrawn, can affect a very large number of orders on the line, and the unacceptable impact requires a manual review prior to going online. We default on Agent to guarantee 100% accuracy, critical operational failure to accept consequences, and human confirmation. AI pages every online version of it for red-line tests to see if it's wrong. Some scenarios could be relaxed, such as an error in extracting amounts, and could be accessed to a manageable extent by measuring the rate of loss. In the end, it was decided whether release or human confirmation would be possible, depending on whether the error was made, the impact was acceptable, the scope and the damage.
Ludto: We're very much like Mr. Yi. For example, the date of the order is wrong, the date is on Friday, we push it on Saturday, the date is not visible on the guide card, the user may pay directly and also use coupons, and then the benefit is gone, the price of the hotel and traffic is changed quickly and the price is gone, so be careful. And the second question is whether the wrong user will find, for example, that the itinerary is clearly unreasonable, that it is going to go back and forth to the same spot, or that it is recommended that the "A" hotel hangs out of the "B" card.
Wang Mo: So toC is more concerned with user experience, toB is using security and responsibility as a red line, who's going to blame for the error, and what's going on with Agent will affect my performance.
Audience: How does Agent optimise the trajectories in the business scene?
Your Majesty: Two. Optimizing the business landscape, now is looking for a badcase analysis to determine whether the intent is wrong, the tool is wrong, or the tool is wrong, but the effect or format of the tool is wrong: changing the prompt or the model if the intent is wrong, the tool call error is an assessment of whether the description is wrong or the tool is too many to be deleted, and it is clear that the combination of either changing the model or changing the Harness is different. The other one we do is the MAS platform, the AI observation and the AI stability, which is still being explored. Now, 70% of the AI agent optimization is still manual, and the industry has some sort of automatic ways of optimising through a trace of naturals or skills, which, frankly, is not working well in the production scenes, but it's working well.
The generic scenario is, for example, OpenClaw, which is very hot, moves and does an internal version for employees, and the departments create a variety of skills to process files and complaints. Harness, who had previously taken the open source, changed the accuracy of the change, the accuracy of leadership requirements, the absence of complaints and the evaluation criteria, so we wanted to use the call chain to assess, for example, skills over 20 steps, and to allow LLM to automatically analyse whether skill was not good or prompt was not good, and optimise dynamically. The utility of this generalized scenario is to reduce the number of steps from more than 20 paces to an average of 15 paces, on average, an increase of about a quarter of efficiency; the accuracy is not assessed well, depending on whether the repetition rate has decreased. So the one-on-one business scenario is people-to-people-to-test set, and the generic scenario is automating.
Rongji: Our experience is two closed rings. The outer ring is a loop for operational indicators. For example, AI sets up the order pairs, which are numerous and each with a transformation funnel, with a system of business indicators to ensure that the right issues are resolved: First find the right driver, then call out using AI, talk to SOP, follow instructions, Agent's ability to talk to the user, and finally make a single; there's also the touch rate, access rate of the outside call line. The outer ring is to make sure that you can find the missing link and the right direction to do your work, that you can't get up there, that you can't get up there, and that you can't get anywhere. The inner ring is the closed circle of Agent's own capabilities. AI calls are open and rely on an observational recording of the whole process to see whether each link is intended to be identified, ASRs, answers are correct, often manually marked, and industry accelerates with LLM as a Judge, or with a more advanced model, saves a lot of manpower. LLM can also be used to extract the words and tactics of the gold medal and accelerate the extraction of business knowledge. So it's called "Agent," and it's both closed: the business indicator system is set, and it's clear which direction Agent works, and Agent's own ability is not followed in words and business strategy, and the badcase rate is down.
Lüdteau: First of all, the evaluation must be accurate, and we need to work hard on the evaluation, because we have some on-line entry criteria that are based on strong codes, some on value models, others that are not. The second is to prepare a data set of gold on the line, called Metadata, which is retrometrically measured every time, to ensure that changes are correct, because self-evolution is easy to hack and models are easy to slack.
Q: The core is to have a learning goal. I need to know that I learned the logic of gold dress, gold sales, where this tag is; and how and how this tag will work. This is the kind of feedback logic that evolves from time to time, including what Mr. Lu said about using the gold standard for offline. By defining the problem first, the language is rich, and Agent ' s self-learning feedback capacity is becoming stronger and linked to online indicators. I've done this in the past, and it's very effective: offline digs find that when people ask for preferential terms, gold-carded customers talk about how to use preferential, re-sell him to influence their decision-making; because you already have a lot of transformational and unconverted outcomes, and it's often people who ask what to answer, what to do, what to do, and how low the conversion rate is.
How much autonomy should Argentina have?
Wang Moo: Now the industry likes to talk about Autonomous Age, as if the more it can plan, mobilize tools, and perform long tasks, the more advanced it is. But in the production environment, the real business of Agent, the more autonomous, the better?
I have two examples. The first is the list, and we've tried to complete it by giving the model context, the target, the tools that can be called, so that it can decide how to interact with the user, and the result is worse than the original process. Because with a clear business scenario, AI will be able to output a lot of content, ask questions for a few seconds, read and understand for a long time; it gives users too much choice when it's free, but a clear user just wants to save himself, and you just tell him what car to pick and at what price. You're using AI yourself to do the PPT just to make it quick, not to discuss it for half a day. The second is AI Schedule, which is also anticipating, leaving the target, context, and the tools to itself to decide how to talk to the user and the driver. The problem is that the path is uncontrollable: the last thing that works is that it doesn't know what strategy it uses, it can't continue to be optimized, and the new strategy can't be followed to recover effects, from strategy to implementation to results. Key nodes therefore require certainty and autonomy only at the point where the model needs to be autonomous.
Ludtau: Agent in the production environment needs to have clear boundaries and responsibilities, more than 99%, if not 100%, or the C-end. It can be flexible in the decision-making space, for example, by calling on tools; but be careful to respond by exporting to the C-end, and it is impossible to get tickets, real-time prices, inventories, commodity IDs, orders in the wrong or wrong or bad.
His Majesty: What is the difference between autonomy and control? I understand that autonomy maximizes its potential under limited rules and that it is out of line. Why are we talking about autonomy? If the mission was a work-like one, it would also require Agent, but it would never be needed to become more autonomous; autonomy was used to solve problems that could not be solved. So the upper limit of autonomy is the degree of flexibility that you want it to solve, but with a threshold, without a threshold, it's out of control. The threshold is technically simple: allow you to call the tools, but it's up to me to do what you can, and your autonomy is just to choose which one you choose. The threshold is 10 or 20 tools.
There's also a problem here: the original programming is called credible programming, and now the generic programming is called untrustworthy programming, because it writes code: gives it 10 tools, it doesn't think it's done, but it's done hard, and it's done by writing it, and it's out of control, so we're all going to make sandboxes. Besides, why is autonomy not good enough? Because 99% of the manufacturers do not have the capacity to upgrade on the basis of their behaviour or error: there are few upgrades to agentic, complicated, or even impossible to find, because most business scenarios are too expensive to recover data for intensive learning. In the future, if an enterprise were able to do so, autonomy could grow, otherwise it would have been limited to the threshold or out of control.
Simulation: Optimistic links to the Enterprise Agent are not sufficiently standardized compared to systems that have developed mature feedback loops, such as recommendations, advertisements, etc. Although it is already possible to use trajectory analysis, offline assessment and feedback data over time, it is still difficult and lacking a broadly reusable production paradigm to achieve a continuous increase in attributable and controlled risks.
Looking ahead
Agent, what is core competitiveness?
亓Longki: The model is the starting point, the base model is open to all, and a good model is higher. At the operational level, the first step is Harnesse, which uses engineering to distinguish between defined and uncertain boundaries: a model that needs to be sent to think about planning, a model that needs to be sent to engineering, and a tool that integrates machine learning algorithms and optimization, which is the lower limit for the company Agent. Above the lower limit, there is a continuous upgrading of the closed loops by operations loop and Agent capacity loop. The upper limit, like engineering, is not the case for a business system, and ultimately not for infrastructure, but for the extraction of business knowledge and the design of the business architecture, which determines whether the system can support operations for five to ten years or whether it has to be substantially restructured each year. By analogy with Agent, the ceiling is on how much business knowledge can be extracted, what scenes to find and where to work. So the model decides the starting point, Harness decides the bottom limit, double Loop decides the direction, and the extent to which business knowledge is extracted and where it works.
King's boss: The Enterprise Agent, I built one of my own, or I built one of my own. The core competitiveness and barriers are different. For us, it's a big challenge to fast-track the ability of an enterprise, and we've done the MCP, Skills (which may not be the type of agreement), AGUI, but it's hard to get stock systems into Agent. I agree very much with the data extraction, and it doesn't make sense to have no data, no data extraction business, Agent. The biggest problem with the production landscape of enterprises is that most of the rules and operational data are deposited in the old system. We've all done some services, like moving, and Chinese oil, and it's hard for you to diagnose a business because you can't get it or get it. So how to combo and expose the capabilities of enterprises that are not well informed in a relatively universal way and become a platform for rapid access is key to the future.
It's been a long time since the hot Palantir model of data business, but it's not going to be able to release it, clean up the data, have a platform, and not use it until LLM is available. What is lacking in most enterprises is a Palantir-type platform that can automate data assets. Also, the key within the enterprise is how to balance management with AI agent itself: whether the enterprise Agent can carry duties and processes. We are not only doing business now, Agent, but also linking AI agent to the original business processes, which correspond to the original responsibilities, such as smart approvals, clearances that require AI agent to match the process to ensure clarity of responsibility and permission. So, on the one hand, the ability to extract enterprise capacity data, and on the other hand, the ability to link Agent and processes are built on sufficiently strong Harness, Loop and data capabilities.
Ludtau: Travel is a relatively lean industry. I think whether the core competitiveness of the future firm Agent, or the supply of industry and industry knowledge in niche areas, can become an enforceable, verifiable and iterative system capacity. For example, when planning a trip, a modulation model can produce a good trip, but the difficulty is whether it really works: the distance between the attractions, the traffic, the business hours, the intensity of the day-to-day play, the attractions actually exist in reality. The journey goes down to buying hotels and tickets, and the inventory of prices must come from the real business system, the tickets must match the real travel needs, and the purchase orders must be confirmed by the users, none of which can be generated by the free exercise of the model. Models are suitable for understanding preferences, screening sites, generating programmes, map route business time can be obtained using tools and ticket combinations can be used as certainty algorithms, but prices and inventories must be provided by the business system. So, instead of generating an apparently good answer, the barriers to business Agent are whether the firm’s own data, tools, algorithms, business rules, and evaluation systems can be set up in the real world.
Wang imitation: It's kind of like a world model, working in the real world. Over the past two years, we have been discussing how to get Agent to be smarter and do more; but after Agent actually went into production, another question became more and more important: what exactly do we allow Agent to do and what must be done by people?
Thermal continuity: December AICon Beijing station
On December 18-19, the Aicon Global Congress on Artificial Intelligence Development and Applications will be held in Beijing. The Congress, which focused on exploring the A application boundary, set up 13 major technical topics covering the core directions of LLM post-training, reasoning optimization, Agent engineering, intellectual engineering, AI agent assessment and security, self-evolution systems, AI software delivery, native architecture and organizational change, bringing together first-line technical experts to share in depth the real practices of AI technological breakthroughs and enterprise size.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.